Large model optimization-based ship main and auxiliary power real-time switching method and system

Abstract

The invention discloses a ship main power and auxiliary power real-time switching method and system based on large model optimization, and relates to the technical field of ship power system control, and the method comprises the steps: collecting the operation state data and environment parameters of a ship main power system and an auxiliary power system in real time through a multi-mode sensor array; and performing space-time alignment and noise suppression processing on the operation state data and the environment parameters, inputting the feature subset into a lightweight large model inference engine, generating a multi-target optimization decision set, executing a main and auxiliary power dynamic switching control sequence in a preset time window according to the multi-target optimization decision set, and outputting the main and auxiliary power dynamic switching control sequence. And generating a self-optimization log including a switching efficiency evaluation report and a model iteration suggestion based on the deviation degree of the switching feedback data and a preset performance index. According to the ship main and auxiliary power real-time switching method and system based on large model optimization, stable operation of a ship power system is guaranteed, and meanwhile the service life of equipment is prolonged.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ship power system control, in particular to a ship main and auxiliary power real-time switching method and system based on large model optimization. BACKGROUND

[0002] The existing ship main and auxiliary power switching technology mostly adopts a fixed threshold triggering mechanism and relies on single sensor data for decision-making, which is difficult to cope with dynamic changes in complex sea conditions. The traditional method has problems such as discrete sampling and noise interference in the data acquisition and processing link, resulting in inaccurate switching time judgment and easy power impact and equipment damage. In addition, the existing system lacks the ability to fuse and analyze multi-source heterogeneous data, and cannot realize comprehensive evaluation of the power system state, which restricts the reliability of the switching decision.

[0003] With the development of artificial intelligence technology, some research attempts to introduce machine learning algorithms to optimize the switching process, but still faces challenges such as insufficient model generalization ability and large computational resource consumption. In particular, in extreme working conditions such as severe sea conditions and sudden load changes, the response speed and decision accuracy of the existing intelligent switching system decrease significantly. At the same time, the traditional method lacks an effective self-learning and optimization mechanism, and is difficult to adapt to the individual needs and long-term performance degradation of different ships, so there is an urgent need for a high-precision, self-adaptive, and interpretable power switching scheme. SUMMARY

[0004] In order to overcome the deficiencies of the above-mentioned prior art, the present application provides a ship main and auxiliary power real-time switching method and system based on large model optimization.

[0005] In a first aspect, the present application provides a ship main and auxiliary power real-time switching method based on large model optimization, the method comprising: real-time acquisition of the running state data and environmental parameters of the ship main and auxiliary power system through a multi-modal sensor array; spatial and temporal alignment and noise suppression processing of the running state data and environmental parameters, construction of an input data set including spatial and temporal coding labels, multi-source data correlation matrices, and abnormal feature vectors, and selection of a feature subset strongly related to power switching decisions through a dynamic feature selection algorithm; inputting the feature subset into a lightweight large model inference engine, extracting power system degradation features and switching risk probability distribution based on a spatial and temporal attention mechanism, and generating a multi-objective optimization decision set including switching time suggestions, power smoothing strategies, and fault suppression schemes; According to the multi-objective optimization decision set, a main and auxiliary power dynamic switching control sequence is executed within a preset time window, and electrical parameters and mechanical vibration indicators during the switching process are fed back in real time; Based on the deviation degree of the switching feedback data and the preset performance index, the parameter weight of the lightweight large model is updated through the federated learning framework, and a self-optimization log including the switching performance evaluation report and the model iteration suggestion is generated.

[0006] Preferably, a multi-source data conflict resolution model is constructed, and when the sensor data is contradictory, the weight is dynamically allocated based on the historical reliability score and the real-time confidence; An outlier detection algorithm is used to remove abnormal data points, and temporal and spatial interpolation is used to repair missing data; The correlation index of the feature and the switching decision is calculated by mutual information entropy, and the feature dimension with a correlation higher than a preset threshold is retained; The filtered feature subset is subjected to adversarial sample detection, and noise is injected to enhance the robustness of the model; When the proportion of redundant features exceeds a preset proportion, a feature space dimension reduction operation is triggered.

[0007] Preferably, a knowledge distillation technique is used to compress a hundred-billion-parameter general large model into a lightweight model suitable for ship edge computing devices, while retaining the core reasoning ability in the field of ship power switching; In the model compression process, a dynamic pruning mechanism is embedded, and the number of model layers and attention heads is dynamically adjusted according to real-time computing resources and task priorities; A ship power digital twin model is constructed, and training data is generated by simulating switching scenarios in different sea conditions to enhance the generalization ability of the model to extreme working conditions; In the model inference stage, a cache optimization strategy is introduced, when the key-value cache exceeds a threshold, a clustering algorithm is used to group and quantize the attention head output, and a cache replacement strategy is used to retain high information entropy features; A multi-granularity inference mode is constructed, and different calculation precisions are selected according to the task urgency: low-precision fast inference is used for real-time switching decision, and high-precision offline calculation is used for post-analysis tasks.

[0008] Preferably, in the pre-synchronization phase matching stage, the amplitude, frequency and phase of the auxiliary power output voltage are adjusted by the virtual synchronous machine technology, so that the error between the main power bus parameters is less than the preset threshold; In the hybrid energy storage buffer energy injection stage, the energy distribution ratio of supercapacitor and flywheel energy storage is calculated according to the power mutation gradient, and millisecond-level power compensation is realized through a bidirectional converter; In the virtual impedance adaptive adjustment stage, the virtual impedance value is dynamically adjusted based on the real-time grid impedance spectrum analysis result, to suppress the circulating current and harmonic oscillation at the switching moment; A multi-objective game algorithm is used to optimize the switching timing, balance the equipment life loss, fuel economy and emission constraints, and generate a Pareto optimal switching path; The reverse check mechanism is started after the switching is completed, and the secondary optimization or alarm signal is triggered by comparing the deviation degree of the actual switching effect and the predicted decision.

[0009] Preferably, a ship power switching knowledge graph is established, historical fault cases, maintenance strategies and expert experience rules are integrated, and the real-time decision is supported by the explainable support provided by the enhanced generation technology through retrieval; A model inference sandbox is deployed on the edge computing device, differential privacy encryption processing is performed on the input data, and the data traceability information of the key decision nodes is stored on the blockchain; When a new fault mode is detected, the active learning mechanism is started to collect sample data and trigger incremental training of the model, and the updated model parameters are synchronized to other ships in the fleet through a secure multi-party computing protocol.

[0010] Preferably, entity relationship triples are extracted from the ship maintenance manual, fault log and operation specification to construct an ontology library covering device topology, fault causal chain and maintenance operation constraints; A graph neural network is used to dynamically embed and represent the knowledge graph, capturing high-order semantic associations between nodes; In the real-time decision-making process, historical cases similar to the current working condition are retrieved through subgraph matching technology, and a decision explanation report including confidence score and risk prompt is generated; A knowledge distillation channel is constructed to convert expert experience rules into a loss function constraint that the model can understand, ensuring that AI decisions comply with industry safety standards.

[0011] Preferably, a multi-modal human-computer interaction interface is constructed, a three-dimensional energy flow topology animation of the power switching process is superimposed on the electronic chart, and the decision logic and risk warning are fed back through natural language interaction; When an uncontrollable abnormality occurs during the switching process, a degradation control mode is started to prioritize the minimum feature set of the propulsion system and autonomously generate an emergency navigation path to avoid high-risk sea areas.

[0012] Preferably, based on the reinforcement learning algorithm, the optimal obstacle avoidance path is calculated within a preset time, taking into account the fuel reserves, sea current direction and nearby rescue resource distribution; The impact range of the power system degradation is marked on the electronic chart interface, including the propulsion force decay curve, the steering response delay threshold and the estimated range; The three-dimensional dissection view of the power switching fault is displayed to the crew through augmented reality equipment, and the key components that require manual intervention are highlighted.

[0013] Preferably, a ship power health index model is constructed to comprehensively evaluate the remaining life, performance degradation rate and maintenance cost weight of the main and auxiliary power systems. When the health index is lower than the preset warning value, a preventive maintenance plan is automatically generated, and a maintenance window is reserved in linkage with a port scheduling system; A virtual pressure test scenario is embedded in the maintenance plan, and the promotion effect of different maintenance schemes on the power switching performance is simulated through digital twinning.

[0014] Preferably, the operating state data includes main diesel engine vibration spectrum, auxiliary gas turbine speed fluctuation, power grid harmonic distortion rate, fuel pressure gradient and cooling water temperature change rate, and the environmental parameters include real-time sea state level, ship roll angular velocity and load mutation prediction value.

[0015] In a second aspect, a ship main and auxiliary power real-time switching system based on large model optimization includes: A data acquisition unit is configured to acquire real-time operating state data and environmental parameters of a ship main and auxiliary power system through a multi-modal sensor array, the operating state data including main diesel engine vibration spectrum, auxiliary gas turbine speed fluctuation, power grid harmonic distortion rate, fuel pressure gradient and cooling water temperature change rate, and the environmental parameters including real-time sea state level, ship roll angular velocity and load mutation prediction value. A data set construction unit is configured to perform spatio-temporal alignment and noise suppression processing on the operating state data and environmental parameters, construct an input data set including spatio-temporal coding labels, multi-source data correlation matrices and abnormal feature vectors, and select a feature subset strongly related to power switching decision through a dynamic feature selection algorithm. A decision set generation unit is configured to input the feature subset into a lightweight large model inference engine, extract power system degradation features and switching risk probability distribution based on a spatio-temporal attention mechanism, and generate a multi-objective optimization decision set including switching time suggestion, power smoothing strategy and fault suppression scheme. A control sequence switching unit is configured to execute a main and auxiliary power dynamic switching control sequence within a preset time window according to the multi-objective optimization decision set, the control sequence including three stages of pre-synchronization phase matching, hybrid energy storage buffer energy injection and virtual impedance adaptive adjustment, and real-time feedback of electrical parameters and mechanical vibration indicators during the switching process. A log generation unit is configured to update the parameter weights of the lightweight large model through a federated learning framework based on the deviation degree of the switching feedback data and the preset performance indicators, and generate a self-optimization log including a switching performance evaluation report and a model iteration suggestion.

[0016] Compared with the prior art, the present application has the following characteristics and beneficial effects: Firstly, the running state data and environmental parameters of the main and auxiliary power systems are collected in real time by a multi-modal sensor array, realizing comprehensive perception of the running state of the ship power system, solving the problem of single data collection dimension in traditional methods, providing multi-dimensional data support for subsequent intelligent decision-making, and significantly improving the comprehensiveness and accuracy of system state monitoring. Secondly, the collected running state data and environmental parameters are processed for time and space alignment and noise suppression, an input data set containing time and space encoding labels, multi-source data correlation matrix and abnormal feature vector is constructed, the time and space deviation and environmental noise interference between sensor data are effectively eliminated, and the most relevant feature subset for power switching decision is selected through a dynamic feature selection algorithm, which not only ensures data quality but also improves the efficiency of subsequent model processing. Further, the feature subset is input into a lightweight large model inference engine for intelligent analysis, the degradation features of the power system and the switching risk probability distribution are extracted based on the space-time attention mechanism, a decision set containing multiple optimization schemes is generated, breaking through the limitations of traditional threshold judgment, realizing adaptive processing capability for complex working conditions, and significantly improving the scientificity and reliability of switching decision. According to the multi-objective optimization decision set, the main and auxiliary power dynamic switching control sequence is executed, and smooth switching is realized through phased fine control strategy, effectively reducing the mechanical impact and electrical disturbance generated in the traditional switching process, ensuring the stable operation of the ship power system, and prolonging the service life of the equipment. Finally, based on the switching feedback data, the model parameters are continuously optimized through the federated learning framework, so that the system can continuously adapt to the changes of ship performance and environmental conditions, realizing the continuous self-improvement and performance improvement of the power switching system. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 is a step block diagram of a ship main and auxiliary power real-time switching method based on large model optimization mainly embodied in this embodiment.

[0018] Figure 2 is a structure block diagram of a ship main and auxiliary power real-time switching system based on large model optimization mainly embodied in this embodiment. DETAILED DESCRIPTION

[0019] The present application will be further described in detail below in conjunction with the following embodiments.

[0020] Reference Figure 1 A ship main and auxiliary power real-time switching method based on large model optimization, the method comprising the following steps: S1. Collecting the running state data and environmental parameters of the ship main and auxiliary power systems in real time by a multi-modal sensor array.

[0021] S2. The operation state data and environmental parameters are subjected to spatio-temporal alignment and noise suppression processing to construct an input data set including spatio-temporal coding labels, multi-source data correlation matrix and abnormal feature vector, and a dynamic feature selection algorithm is used to screen a feature subset strongly related to power switching decision.

[0022] S3. The feature subset is input into a lightweight large model inference engine to extract power system degradation features and switching risk probability distribution based on a spatio-temporal attention mechanism, and a multi-objective optimization decision set including switching timing suggestion, power smoothing strategy and fault suppression scheme is generated.

[0023] S4. According to the multi-objective optimization decision set, a main and auxiliary power dynamic switching control sequence is executed within a preset time window, and electrical parameters and mechanical vibration indicators during switching are fed back in real time.

[0024] S5. Based on the deviation degree of switching feedback data and preset performance indicators, the parameter weights of the lightweight large model are updated through a federated learning framework, and a self-optimization log including switching performance evaluation report and model iteration suggestion is generated.

[0025] Specifically, first, the operation state data and environmental parameters of the main and auxiliary power systems are collected in real time by a multi-modal sensor array, realizing comprehensive perception of the ship power system operation state, solving the problem of single data collection dimension in traditional methods, providing multi-dimensional data support for subsequent intelligent decision-making, and significantly improving the comprehensiveness and accuracy of system state monitoring. Secondly, the collected operation state data and environmental parameters are subjected to spatio-temporal alignment and noise suppression processing to construct an input data set containing spatio-temporal coding labels, multi-source data correlation matrix and abnormal feature vector, effectively eliminating the spatio-temporal deviation and environmental noise interference between sensor data, and selecting the most relevant feature subset for power switching decision through a dynamic feature selection algorithm, which not only guarantees data quality but also improves the efficiency of subsequent model processing. Further, the feature subset is input into a lightweight large model inference engine for intelligent analysis, and the power system degradation features and switching risk probability distribution are extracted based on a spatio-temporal attention mechanism to generate a decision set containing multiple optimization schemes, breaking through the limitations of traditional threshold judgment, realizing adaptive processing capability for complex working conditions, and significantly improving the scientificity and reliability of switching decision. According to the multi-objective optimization decision set, a main and auxiliary power dynamic switching control sequence is executed, and smooth switching is achieved through phased fine control strategy, effectively reducing mechanical impact and electrical disturbance generated in traditional switching process, ensuring stable operation of the ship power system, and prolonging the service life of the equipment. Finally, based on the switching feedback data, the model parameters are continuously optimized through a federated learning framework, so that the system can continuously adapt to changes in ship performance and environmental conditions, realizing continuous self-improvement and performance improvement of the power switching system.

[0026] The specific process of step S1 can be that the system adopts a distributed sensor network architecture, high-precision vibration sensors are installed at key positions of the main diesel engine and the auxiliary gas turbine, the sampling frequency is not less than 10 kHz, and the subtle characteristics of mechanical vibration can be captured. The power grid harmonic monitoring adopts a wideband current transformer combined with a high-speed data acquisition card to realize real-time analysis of 0-50 harmonics. The fuel system is configured with a multi-channel pressure sensor array to monitor the pressure gradient from the fuel pump outlet to the fuel nozzle. The cooling water temperature monitoring uses a platinum resistance temperature sensor combined with a digital transmitter, with a temperature resolution of 0.1°C. The environmental parameter acquisition module integrates ship motion attitude sensors, weather radars and sea state analyzers, and realizes accurate estimation of ship motion parameters through Kalman filtering algorithm. All sensor data are synchronized to the microsecond level through time synchronization protocol, providing an accurate time reference for subsequent analysis. This multi-source data acquisition scheme breaks through the limitations of traditional single sensor monitoring and realizes stereoscopic perception of the running state of the power system.

[0027] The specific step S2 includes the following sub-steps: A multi-source data conflict resolution model is constructed, and when the sensor data is contradictory, the weight is dynamically allocated based on the historical reliability score and real-time confidence; An outlier detection algorithm is used to remove abnormal data points and perform time and space interpolation repair on missing data; The correlation index between features and switching decisions is calculated through mutual information entropy, and the feature dimensions with a correlation higher than a preset threshold are retained; The filtered feature subset is subjected to adversarial sample detection to inject noise and enhance the robustness of the model; When the proportion of redundant features exceeds a preset proportion, a feature space dimension reduction operation is triggered.

[0028] Specifically, the data processing module first uses a sliding window correlation analysis method to time-align the multi-source data, eliminating the time deviation caused by transmission delay. Spatial alignment is achieved by establishing a unified ship coordinate system, and all sensor data are converted to this coordinate system for analysis. Noise suppression uses wavelet transform combined with an adaptive threshold denoising algorithm to effectively retain signal features while suppressing random noise. In the feature extraction stage, a deep convolutional neural network is used to automatically learn the time-frequency features of the signal and generate a high-dimensional feature vector. Dynamic feature selection calculates feature importance scores based on the random forest algorithm to real-time filter the feature combination that best discriminates the current working condition. This processing procedure significantly improves data quality and provides a reliable input for subsequent intelligent decision-making.

[0029] The specific step S3 includes the following sub-steps: The knowledge distillation technology is used to compress the hundred-billion-parameter general large model into a lightweight model suitable for ship edge computing devices, while retaining the core reasoning ability in the field of ship power switching. A dynamic pruning mechanism is embedded in the model compression process to dynamically adjust the number of model layers and attention heads according to real-time computing resources and task priorities. A ship power digital twin model is constructed to simulate switching scenarios in different sea conditions and generate training data, enhancing the model's generalization ability in extreme conditions. In the model inference stage, a cache optimization strategy is introduced. When the key-value cache exceeds the threshold, a clustering algorithm is used to group and quantize the attention head output, and a cache replacement strategy is used to retain high information entropy features. A multi-granularity inference mode is constructed to select different calculation precisions according to the urgency of the task: low-precision fast inference for real-time switching decisions, and high-precision offline calculation for post-analysis tasks.

[0030] Specifically, the lightweight large model can use the Transformer architecture, and through the knowledge distillation technology, the parameter quantity of the original model can be compressed by more than 90%, while retaining more than 95% of the prediction accuracy. The model input layer designs a spatiotemporal attention mechanism to automatically learn the relevance of different sensor data in the time and space dimensions. The middle layer uses residual connection and layer normalization technology to ensure the stability of the deep network training. The output layer uses a multi-task learning framework to simultaneously predict the switching time, power compensation demand, and fault risk, generating comprehensive decision recommendations. The inference process uses mixed precision calculation, reducing the calculation time by 40% while ensuring accuracy. This intelligent decision model realizes accurate judgment of complex conditions and significantly improves the scientific nature of switching decisions.

[0031] The specific step S4 includes the following sub-steps: In the pre-synchronization phase matching stage, the amplitude, frequency, and phase of the auxiliary power output voltage are adjusted through the virtual synchronous machine technology to make the error between the main power bus parameters and the preset threshold less than the preset threshold; In the hybrid energy storage buffer energy injection stage, the energy distribution ratio of supercapacitor and flywheel energy storage is calculated according to the power mutation gradient, and millisecond-level power compensation is realized through a bidirectional converter; In the virtual impedance adaptive adjustment stage, the virtual impedance value is dynamically adjusted based on the real-time grid impedance spectrum analysis results to suppress the circulating current and harmonic oscillation during switching; A multi-objective game algorithm is used to optimize the switching timing, balancing device life loss, fuel economy, and emission constraints to generate a Pareto optimal switching path; After the switching is completed, a reverse verification mechanism is started to compare the deviation between the actual switching effect and the predicted decision, triggering a secondary optimization or alarm signal.

[0032] Specifically, the pre-synchronization stage adopts a precise phase control algorithm based on a phase-locked loop to control the phase difference of the auxiliary power output voltage within 0.5 degrees. The energy buffer stage dynamically adjusts the power distribution ratio of the super capacitor and the flywheel energy storage through model predictive control to achieve millisecond-level power compensation. The impedance adjustment stage uses a fuzzy PID algorithm to optimize the virtual impedance parameters in real time based on online impedance measurement results. The entire switching process uses an event-triggered mechanism, which immediately starts the protection strategy when abnormal vibration or electrical parameter out-of-limit is detected. This control scheme realizes seamless switching between the main and auxiliary power, and controls the power fluctuation during switching within 5%.

[0033] Specifically, the process of step S5 can be that the federated learning client is deployed on each ship edge computing node, and the model update gradient is uploaded to the cloud aggregation server after encryption. The server aggregates the node gradients using a secure multi-party computation protocol to generate a global model update. The model evaluation module quantifies the performance indicators of each switching from the switching smoothness, energy efficiency and equipment loss dimensions to generate a detailed evaluation report. When the model performance is detected to be degraded, the incremental training process is automatically triggered to optimize the model parameters using the latest data. This continuous learning mechanism ensures that the system can adapt to changing operating environments.

[0034] In some embodiments, the ship main and auxiliary power real-time switching based on large model optimization can further include the following steps: A ship power switching knowledge graph is established to integrate historical fault cases, maintenance strategies and expert experience rules, and a retrieval enhancement generation technology is used to provide explainability support for real-time decision-making; A model inference sandbox is deployed on an edge computing device to perform differential privacy encryption on input data, and the data traceability information of key decision nodes is stored on a blockchain; When a new fault mode is detected, a proactive learning mechanism is started to collect sample data and trigger incremental training of the model, and the updated model parameters are synchronized to other ships in the fleet through a secure multi-party computation protocol.

[0035] Specifically, the knowledge graph construction uses a Neo4j graph database to store entity relationships, and uses a BERT model to automatically extract the association between fault features and treatment measures from maintenance manuals and fault reports. The graph contains more than 100,000 entity nodes and 500,000 relationship edges, covering all typical fault modes of the ship power system. During real-time decision-making, the system calculates the semantic similarity between the current working condition and historical cases using graph embedding technology, and retrieves the most relevant maintenance experience. The retrieval results are integrated into the reasoning process of the large model through an attention mechanism to enhance the explainability of the decision. At the same time, the system automatically records new fault handling experience to continuously expand the coverage of the knowledge graph. This knowledge enhancement mechanism significantly improves the system's ability to handle new faults.

[0036] In some embodiments, the large model optimization-based real-time switching method of ship main and auxiliary power can further include the following steps: Extracting entity relationship triples from ship maintenance manuals, fault logs and operation specifications to build an ontology library covering equipment topology, fault causal chain and maintenance operation constraints; Using graph neural networks to dynamically embed knowledge graphs to capture high-order semantic associations between nodes; In the real-time decision-making process, historical cases similar to the current working condition are retrieved through subgraph matching technology, and a decision explanation report including confidence score and risk prompt is generated; Build a knowledge distillation channel to convert expert experience rules into model-understandable loss function constraints to ensure that AI decisions comply with industry safety standards.

[0037] Specifically, the edge computing node uses containerization technology to deploy model inference services, and each service runs in an independent sandbox environment to ensure fault isolation. Data privacy protection uses the ε-differential privacy algorithm to add noise conforming to the Laplace distribution in the feature extraction stage. The blockchain network is built based on the Hyperledger Fabric framework, and records key information such as the timestamp, input features and output results of each decision. The smart contract automatically verifies the integrity and consistency of the data to prevent malicious tampering. This security architecture not only protects data privacy, but also ensures the auditable nature of the decision-making process.

[0038] In some embodiments, the large model optimization-based real-time switching method of ship main and auxiliary power can further include the following steps: Build a multi-modal human-computer interaction interface to display a three-dimensional energy flow topology animation of the power switching process on the electronic chart, and provide feedback on decision logic and risk warning through natural language interaction; When uncontrollable abnormalities occur during the switching process, start the degradation control mode to prioritize the minimum feature set of the propulsion system and automatically generate an emergency navigation path to avoid high-risk sea areas.

[0039] Specifically, the human-machine interaction system adopts a layered visualization architecture. The bottom layer integrates an electronic chart engine, the middle layer superimposes a three-dimensional model of the power system, and the upper layer renders real-time energy flow animations. Energy flow visualization uses particle system technology, with particles of different colors and trajectories representing the transmission paths of electrical, mechanical, and thermal energy. The natural language interaction module is based on a pre-trained language model and supports both voice and text input. It can understand complex queries such as "Why is it recommended to switch now?" and generate explanations containing technical parameters. The risk warning uses a multi-level prompt mechanism. Ordinary reminders are displayed in the status bar, and important warnings trigger full-screen flashing and voice broadcasting. The interface layout is optimized through human factors engineering, with key information always within the operator's optimal field of view. This system significantly reduces the cognitive load of operators and makes complex technical decisions intuitive and easy to understand. The degradation control module uses a finite state machine design, defining three operating modes: "normal," "warning," and "serious." When key parameters exceed safety thresholds, the system automatically enters a degraded mode, shutting down non-essential loads while retaining core functions such as propulsion, navigation, and communication. The emergency path planning is based on an improved RRT* algorithm, considering real-time sea conditions, fuel reserves, and nearby port distribution to generate the optimal risk avoidance route within seconds. The system continuously monitors fault development and automatically sends distress signals and starts black box data protection when it determines that the fault cannot be self-recovered. This mechanism provides the last safety guarantee for sudden failures and significantly improves the survivability of ships.

[0040] In some embodiments, the ship main and auxiliary power real-time switching method based on large model optimization can further include the following steps: Based on the reinforcement learning algorithm, the optimal obstacle avoidance path is calculated within a preset time, considering fuel reserves, sea current direction, and nearby rescue resource distribution; Mark the impact range of the power system degradation on the electronic chart interface, including the propulsion force attenuation curve, steering response delay threshold, and estimated range; Show the crew a three-dimensional dissection view of the power switching fault through augmented reality equipment and highlight the key components that require manual intervention.

[0041] Specifically, the path planning algorithm adopts the deep deterministic policy gradient (DDPG) framework, and the training environment simulates various typical sea conditions and fault scenarios. The state space includes ship dynamics parameters, environmental information, and system status, with more than 50 dimensions. The action space is defined as the continuous variation of rudder angle and throttle opening. The reward function carefully designs multiple indicators such as safety, efficiency, and comfort. During online reasoning, the algorithm re-evaluates the path every 5 seconds and dynamically adjusts the obstacle avoidance strategy. The system also maintains a global port database and can intelligently recommend the nearest safe docking point. The success rate of path planning in complex sea conditions is more than 40% higher than traditional methods. The influence range visualization uses contour filling technology, and different color areas represent different performance attenuation levels. The propulsive force curve is calculated in real time by the ship motion mathematical model, considering factors such as current engine power, propeller efficiency, and hull resistance. The turning delay index is based on the response characteristics of the rudder and the ship maneuverability derivative prediction. The estimated range is integrated with fuel consumption rate, sea current assistance / resistance, and other multi-source data. All prediction results are updated every 30 seconds and displayed in both digital labels and trend charts. This intuitive performance prediction helps the crew accurately determine the maneuvering capability boundary of the ship.

[0042] In some embodiments, the large model optimization-based real-time switching method for ship main and auxiliary power can further include the following steps: Construct a ship power health index model to comprehensively evaluate the remaining life, performance degradation rate, and maintenance cost weight of the main and auxiliary power systems; When the health index is lower than the preset warning value, automatically generate a preventive maintenance plan and link with the port scheduling system to reserve a maintenance window; Embed virtual stress test scenarios in the maintenance plan to simulate the improvement effect of different maintenance schemes on power switching performance through digital twinning.

[0043] Specifically, the health degree model adopts a multi-level fuzzy comprehensive evaluation method. The bottom-level indicators include more than 200 sensor parameters, the middle-level evaluation subsystem state, and the top-level output overall health score. The remaining life prediction is based on the Wiener process degradation modeling, considering the workload and environmental stress acceleration factors. The performance degradation analysis adopts the benchmark comparison method, establishing the deviation index from the factory parameters. The cost weight is dynamically adjusted by the expert system, reflecting the current maintenance resource supply and demand situation in the shipping market. The health degree is directly displayed as 0-100 points, with specific maintenance recommendations. The model realizes the transition from "breakdown maintenance" to "predictive maintenance". The maintenance plan generator is based on the constraint satisfaction problem (CSP) framework, considering the overall consideration of spare parts inventory, technical personnel scheduling, and port berth resource constraints. The system interacts with the port management system in real time through the EDI interface, intelligently recommending time windows and maintenance teams. The plan optimization goals include multiple indicators such as the shortest maintenance time and the minimum downtime loss. The generated plan is automatically synchronized to the ship calendar and shore-based supervision system, starting two weeks in advance to count down and remind. In emergency situations, the system can start a green channel to coordinate priority maintenance resources. The intelligent scheduling reduces unplanned downtime by more than 35%. The digital twin system adopts multi-physics field coupling simulation technology to accurately model the interaction of mechanical, electrical, fluid, and other fields. The pressure test scenario library contains more than 50 typical working conditions such as extreme sea conditions and emergency maneuvers. The maintenance scheme evaluation adopts the A / B testing framework, simulating the effect difference of different strategies such as "replace only worn parts" and "full set of preventive maintenance". The simulation results are presented in the form of three-dimensional animation and parameter comparison table, directly showing the expected life extension and failure rate reduction of each scheme. The virtual verification improves the decision-making scientificity of the maintenance scheme to a new level.

[0044] A ship main and auxiliary power real-time switching system based on large model optimization, by applying a ship main and auxiliary power real-time switching method based on large model optimization as described above, comprising a data acquisition unit, a data set construction unit, a decision set generation unit, a control sequence switching unit, and a log generation unit, referring to Figure 2, the operation state data of the main and auxiliary power system of the ship and the environmental parameters are collected in real time by the data acquisition unit through the multi-modal sensor array, the operation state data includes the vibration frequency spectrum of the main diesel engine, the speed fluctuation of the auxiliary gas turbine, the harmonic distortion rate of the power grid, the fuel pressure gradient and the cooling water temperature change rate, the environmental parameters include the real-time sea state level, the ship roll angular velocity and the load sudden change prediction value; the operation state data and the environmental parameters are processed by the data set construction unit for time and space alignment and noise suppression, the input data set including the time and space coding label, the multi-source data correlation matrix and the abnormal feature vector is constructed, and the dynamic feature selection algorithm is used to screen the feature subset strongly related to the power switching decision; the feature subset is input into the lightweight large model reasoning engine by the decision set generation unit, the degradation features of the power system and the switching risk probability distribution are extracted based on the time and space attention mechanism, and the multi-objective optimization decision set including the switching time suggestion, the power smoothing strategy and the fault suppression scheme is generated; the control sequence switching unit executes the main and auxiliary power dynamic switching control sequence in the preset time window according to the multi-objective optimization decision set, the control sequence includes three stages of pre-synchronization phase matching, hybrid energy storage buffer energy injection and virtual impedance adaptive adjustment, and the electrical parameters and mechanical vibration indexes in the switching process are fed back in real time; the log generation unit updates the parameter weight of the lightweight large model through the federal learning framework based on the deviation degree of the switching feedback data and the preset performance index, and generates the self-optimization log including the switching efficiency evaluation report and the model iteration suggestion.

[0045] The above are preferred embodiments of the present application, and do not limit the protection scope of the present application, therefore: any equivalent changes made on the structure, shape, principle of the present application should be covered within the protection scope of the present application.

Claims

1. A large model optimization-based real-time switching method for ship main and auxiliary power, characterized in that, The method comprises the following steps: Real-time acquisition of ship main and auxiliary power system operation state data and environmental parameters through a multi-modal sensor array; Temporal and spatial alignment and noise suppression processing of the operation state data and environmental parameters, construction of an input data set including temporal and spatial coding labels, multi-source data correlation matrices, and abnormal feature vectors, and screening of a feature subset strongly related to power switching decisions through a dynamic feature selection algorithm; Inputting the feature subset into a lightweight large model inference engine, extracting power system degradation features and switching risk probability distribution based on a temporal and spatial attention mechanism, and generating a multi-objective optimization decision set including switching time suggestion, power smoothing strategy, and fault suppression scheme; According to the multi-objective optimization decision set, executing a main and auxiliary power dynamic switching control sequence within a preset time window, and real-time feedback of electrical parameters and mechanical vibration indicators during the switching process; Based on the deviation degree of switching feedback data and preset performance indicators, updating the parameter weights of the lightweight large model through a federated learning framework, and generating a self-optimization log including a switching performance evaluation report and model iteration suggestions.

2. The method according to claim 1, wherein, The steps of temporal and spatial alignment and noise suppression processing of the operation state data and environmental parameters, construction of an input data set including temporal and spatial coding labels, multi-source data correlation matrices, and abnormal feature vectors, and screening of a feature subset strongly related to power switching decisions are as follows: A multi-source data conflict resolution model is constructed. When sensor data conflicts occur, weights are dynamically allocated based on historical reliability scores and real-time confidence; Outlier data points are removed using an outlier detection algorithm, and missing data is repaired using temporal and spatial interpolation; The correlation index of features and switching decisions is calculated through mutual information entropy, and feature dimensions with a correlation higher than a preset threshold are retained; The screened feature subset is subjected to adversarial sample detection to enhance the robustness of the model; When the proportion of redundant features exceeds a preset proportion, a feature space dimension reduction operation is triggered.

3. The method according to claim 2, wherein, The steps of inputting the feature subset into a lightweight large model inference engine, extracting power system degradation features and switching risk probability distribution based on a temporal and spatial attention mechanism, and generating a multi-objective optimization decision set including switching time suggestion, power smoothing strategy, and fault suppression scheme are as follows: A knowledge distillation technique is used to compress a hundred-billion-parameter general large model into a lightweight model suitable for ship edge computing devices, while retaining the core reasoning ability in the field of ship power switching; A dynamic pruning mechanism is embedded in the model compression process to dynamically adjust the number of model layers and attention heads based on real-time computing resources and task priorities; A ship power digital twin model is constructed to generate training data by simulating switching scenarios in different sea conditions, enhancing the model's generalization ability in extreme conditions; In the model inference stage, a cache optimization strategy is introduced. When the key-value cache exceeds a threshold, a clustering algorithm is used to group and quantize the attention head output, and a cache replacement strategy is used to retain high information entropy features; A multi-granularity reasoning mode is constructed to select different calculation precisions according to the urgency of the task: low-precision fast reasoning for real-time switching decisions, and high-precision offline calculation for post-analysis tasks.

4. The method according to claim 3, characterized in that, According to the multi-objective optimization decision set, a main auxiliary power dynamic switching control sequence is executed within a preset time window, and electrical parameters and mechanical vibration indicators during switching are fed back in real time, specifically as follows: In the pre-synchronization phase matching stage, the amplitude, frequency and phase of the auxiliary power output voltage are adjusted through the virtual synchronous machine technology, so that the error between the auxiliary power output voltage and the main power bus parameters is less than a preset threshold; In the hybrid energy storage buffer energy injection stage, the energy distribution ratio of the super capacitor and the flywheel energy storage is calculated according to the power mutation gradient, and the millisecond level power compensation is realized through the bidirectional converter; In the virtual impedance adaptive adjustment stage, the virtual impedance value is dynamically adjusted based on the real-time grid impedance spectrum analysis result to suppress the circulating current and harmonic oscillation at the switching moment; A multi-objective game algorithm is used to optimize the switching time sequence, balance the device life loss, fuel economy and emission constraints, and generate a Pareto optimal switching path; After the switching is completed, a reverse verification mechanism is started, the deviation between the actual switching effect and the predicted decision is compared, and a secondary optimization or alarm signal is triggered.

5. The method of claim 1, wherein, The method further comprises: A ship power switching knowledge graph is established, historical fault cases, maintenance strategies and expert experience rules are integrated, and a retrieval enhanced generation technology is used to provide explainability support for real-time decision making; A model inference sandbox is deployed on an edge computing device, differential privacy encryption processing is performed on input data, and data traceability information of key decision nodes is stored on a blockchain; When a new fault mode is detected, a proactive learning mechanism is started to collect sample data and trigger incremental training of the model, and the updated model parameters are synchronized to other ships in the fleet through a secure multi-party computation protocol.

6. The method according to claim 5, wherein, The method further comprises: Entity relationship triples are extracted from ship maintenance manuals, fault logs and operation specifications to construct an ontology library covering device topology, fault causal chain and maintenance operation constraints; A graph neural network is used to dynamically embed and represent the knowledge graph to capture high-order semantic associations between nodes; In the real-time decision making process, a subgraph matching technology is used to retrieve historical cases similar to the current working condition, and a decision explanation report including confidence scores and risk warnings is generated; A knowledge distillation channel is constructed to convert expert experience rules into a loss function constraint that can be understood by the model, ensuring that AI decisions comply with industry safety standards.

7. The method of claim 1, wherein, The method further comprises: A multi-modal human-machine interaction interface is constructed to display a three-dimensional energy flow topology animation of the power switching process on an electronic chart, and to provide decision logic and risk warnings through natural language interaction; When an uncontrollable abnormality occurs during the switching process, a degraded control mode is started to prioritize the minimum feature set of the propulsion system and to autonomously generate an emergency navigation path to avoid high-risk sea areas.

8. The method according to claim 7, characterized in that, The method further comprises: An entity relationship triples are extracted from ship maintenance manuals, fault logs and operation specifications to construct an ontology library covering device topology, fault causal chain and maintenance operation constraints; A graph neural network is used to dynamically embed and represent the knowledge graph to capture high-order semantic associations between nodes; In the real-time decision making process, a subgraph matching technology is used to retrieve historical cases similar to the current working condition, and a decision explanation report including confidence scores and risk warnings is generated; A knowledge distillation channel is constructed to convert expert experience rules into a loss function constraint that can be understood by the model, ensuring that AI decisions comply with industry safety standards. Based on reinforcement learning algorithm, the optimal obstacle avoidance path is calculated within a preset time, considering fuel reserves, sea current direction and nearby rescue resource distribution; The influence range of power system degradation is marked on the electronic chart interface, including propulsion force attenuation curve, steering response delay threshold and estimated range; Through augmented reality equipment, the crew is shown a three-dimensional dissection view of the power switching failure, and the key components that require manual intervention are highlighted.

9. The method according to claim 1, wherein, The method further comprises: A ship power health index model is constructed to comprehensively evaluate the remaining life, performance degradation rate and maintenance cost weight of the main and auxiliary power systems; When the health index is lower than the preset warning value, a preventive maintenance plan is automatically generated and a maintenance window is reserved in conjunction with the port scheduling system; A virtual stress test scenario is embedded in the maintenance plan to simulate the improvement effect of different maintenance schemes on power switching performance through digital twinning.

10. The method of claim 1, wherein, The operating state data includes main diesel engine vibration spectrum, auxiliary gas turbine speed fluctuation, power grid harmonic distortion rate, fuel pressure gradient and cooling water temperature change rate, and the environmental parameters include real-time sea state level, ship roll angular velocity and load mutation prediction value.

11. A large model optimization-based ship main and auxiliary power real-time switching system, characterized in that, The system is used to implement the ship main and auxiliary power real-time switching method based on large model optimization according to any one of claims 1-10, comprising: A data acquisition unit is used to acquire real-time operating state data and environmental parameters of the ship main and auxiliary power system through a multi-modal sensor array; A data set construction unit is used to perform time-space alignment and noise suppression processing on the operating state data and environmental parameters, construct an input data set including time-space coding labels, multi-source data association matrices and abnormal feature vectors, and select a feature subset strongly related to power switching decision through a dynamic feature selection algorithm; A decision set generation unit is used to input the feature subset into a lightweight large model reasoning engine, extract power system degradation features and switching risk probability distribution based on a time-space attention mechanism, and generate a multi-objective optimization decision set including switching time suggestion, power smoothing strategy and fault suppression scheme; A control sequence switching unit is used to execute the main and auxiliary power dynamic switching control sequence within a preset time window according to the multi-objective optimization decision set, and real-time feedback electrical parameters and mechanical vibration indicators during switching; A log generation unit is used to update the parameter weight of the lightweight large model through a federated learning framework based on the deviation of switching feedback data and preset performance indicators, and generate a self-optimization log including switching efficiency evaluation report and model iteration suggestion.

Images